Ph.D. Student. Open-source enthusiast. Computer Architecture and software co-design. Intern at $INTC. Opinions are my own. GH: https://xitter.cf/t.co/QtjKsKmkNE

Joined May 2014
大家好,今天我刚上飞机,听到旁边朋友说“飞机从350换到了777,没有Wi-Fi”。为了飞机上继续改 Slides,我马上拿出电脑,git pull 一下 llama.cpp,并编译看看本地已有的 Qwen 3.8 27B + dflash2 能不能跑更快,并迅速安装 MTPLX,马上开始下载模型,发现5G热点还能跑 20MB/s,1000 秒可以下载完。最后这场闹剧结束于“这架飞机有Wi-Fi”,谢谢大家!
7
88
13,974
然后我今天意识到:我的 M4 Pro 跑 MTPLX 的 Qwen 3.8 27B 的推理速度,也就大概和 Astra 的 API 一样慢。
49
I am no longer surprised by dataflow processor architecture papers that have no compiler implementation but run generic software workloads. Because LLMs can act as such compilers for this task today.
1
209
很可惜的是 AirChina 这个 WiFi 不给 SSH、不给 UDP,甚至 443 端口证书不对都要被 DPI 直接 TCP Reset,让我没法进内网连接我的任何服务器,以及任何我的私有部署 LLM。最后在朋友 @sohasaikou 的帮助下给了我一个 websocket over https 才真正连上一个能工作的网。
1
8
1,673
今日尝试:自己用 SGP.26 发 eSIM ,自己用 Open5GS 建LTE 核心网,自己建 LTE RAN ,自己建 IMS 并成功让两台手机打电话和联网
78
175
33
1,997
162,459
For someone who wanna try without RAN, I have good news: with an iPhone (iOS 27.0.1), we can still connect to VoWiFi without RAN by setting our DNS hosts to redirect the ePDG domain to our VoWiFi IPSec IP. Authentication uses EAP-AKA, so we don’t need to trust a self-signed root CA on our phone! I haven’t tried Wi-Fi Calling over Cellular, but iOS allows a profile to configure the DoH server via Apple Configurator, so maybe it will also work.
2
1
5
336
And two sad pieces of news: iPhone will not use EF_ePDGId from the SIM Card. The epdg domain will not be resolved using the DoH profile on iPhone.
94
0!=1 retweeted
For someone who wanna try without RAN, I have good news: with an iPhone (iOS 27.0.1), we can still connect to VoWiFi without RAN by setting our DNS hosts to redirect the ePDG domain to our VoWiFi IPSec IP. Authentication uses EAP-AKA, so we don’t need to trust a self-signed root CA on our phone! I haven’t tried Wi-Fi Calling over Cellular, but iOS allows a profile to configure the DoH server via Apple Configurator, so maybe it will also work.
2
1
5
336
Just a note for people who see this tweet and wanna try to run the LTE / NR RAN: 0) Check local radio regulations. 1) Keep all devices in a faraday cage if anything unsure. 2) NEVER set MCC / MNC to match any existing carrier. 3) Don't use the frequency already allocated for your local carrier or any other existing usage. 4) Use the lowest transmit power configured in the eNB and the UE's transmission control if unsure, even in a faraday cage. And feel free to reach out to me: xitter.cf/yangyuc/status/2110650…
I was astonished that my X follower count doubled from one tweet. That's the research infra for the cellular network that I want to try. I first learned that we can run the core for a cellular network with open-source infrastructure from 1 year ago, when I travelled to ACM CCS 2025 and attended the free5GC workshop. blog.cyyself.name/ccs-2025-t… I'm just a hobbyist in networking (I learned RDMA for HPC. I have my own ASN, my own overlay, etc.), and I want to share my experience with LTE/5G bufferbloat from my blog blog.cyyself.name/5g-nr-tcp-… for them to discuss whether there is any solution. At the workshop, I asked them, "How do you test your 5G core with current UE?", and they answered, "Using UERANSIM". Then I was just wondering: can we set up a RAN in my HomeLab with reasonable cost? Then, I found some cheap LTE femtos on the Chinese second-hand market XianYu (闲鱼), where I can get an LTE B40 femto for ~50 CNY (7.5 USD). I bought one, and I left it alone for a year. In April 2026, I travelled to Edinburgh, UK for EuroSys 2026. My friend @FlyGoat showed me around their lab and their Campus5G network campus5g-guest.netsys-edinbu…, which is a cool thing I really want to try to build in my own lab. Recently, my friend @YukariEwe told me that we can get an SGP.26 eUICC from Taobao with only 80 CNY, so I chose to buy one and set up the Open5GS with my LTE femto using an AI Agent in almost one shot. I really thank these two friends who brought me there. I really want to discover something cool, like the security problem: if we can install an eSIM in one shot now, is the IMS implementation in our phone really hardened enough against the operator-side attack? Also, I want to reproduce the result shown in my blog from 2 years ago: I found TCP BBR is a good solution to cellular bufferbloat; can we solve it in the core network? Mobile Communication isn't my research topic in my PhD life, but I'm curious about many things and want to make friends who can exchange these ideas and problems together. With AI agents now, I can dive into many things with low time cost. So if anyone wants to discuss this or even cooperate with me, feel free to reach out to me!
1
23
4,091
I was astonished that my X follower count doubled from one tweet. That's the research infra for the cellular network that I want to try. I first learned that we can run the core for a cellular network with open-source infrastructure from 1 year ago, when I travelled to ACM CCS 2025 and attended the free5GC workshop. blog.cyyself.name/ccs-2025-t… I'm just a hobbyist in networking (I learned RDMA for HPC. I have my own ASN, my own overlay, etc.), and I want to share my experience with LTE/5G bufferbloat from my blog blog.cyyself.name/5g-nr-tcp-… for them to discuss whether there is any solution. At the workshop, I asked them, "How do you test your 5G core with current UE?", and they answered, "Using UERANSIM". Then I was just wondering: can we set up a RAN in my HomeLab with reasonable cost? Then, I found some cheap LTE femtos on the Chinese second-hand market XianYu (闲鱼), where I can get an LTE B40 femto for ~50 CNY (7.5 USD). I bought one, and I left it alone for a year. In April 2026, I travelled to Edinburgh, UK for EuroSys 2026. My friend @FlyGoat showed me around their lab and their Campus5G network campus5g-guest.netsys-edinbu…, which is a cool thing I really want to try to build in my own lab. Recently, my friend @YukariEwe told me that we can get an SGP.26 eUICC from Taobao with only 80 CNY, so I chose to buy one and set up the Open5GS with my LTE femto using an AI Agent in almost one shot. I really thank these two friends who brought me there. I really want to discover something cool, like the security problem: if we can install an eSIM in one shot now, is the IMS implementation in our phone really hardened enough against the operator-side attack? Also, I want to reproduce the result shown in my blog from 2 years ago: I found TCP BBR is a good solution to cellular bufferbloat; can we solve it in the core network? Mobile Communication isn't my research topic in my PhD life, but I'm curious about many things and want to make friends who can exchange these ideas and problems together. With AI agents now, I can dive into many things with low time cost. So if anyone wants to discuss this or even cooperate with me, feel free to reach out to me!
今日尝试:自己用 SGP.26 发 eSIM ,自己用 Open5GS 建LTE 核心网,自己建 LTE RAN ,自己建 IMS 并成功让两台手机打电话和联网
4
3
1
78
9,691
As my X followers rise rapidly, it's hard to determine which language to use for my X post. Chinese or English.
2
4
640
为什么没有人关注 SGP.26
1
20
10,656
And for the next, I want to explore why this happens: blog.cyyself.name/5g-nr-tcp-…
2
30
11,944
提示:实验请在合法在法拉第笼中完成,虽然 LTE B40 中,三大运营商占用了 2300-2390,最后的 10MHz 处于 Free 状态
2
4
185
16,485
I finally understand why people buy Mac mini for running agents. With macOS, you can run many iOS apps out of the box directly from the App Store. In China, many everyday applications have no web version, or the functionality is downgraded. For example, Amap (高德地图) has no reviews in the web version, and they didn't provide any other client apps for Windows and Linux. So macOS becomes the simplest solution for daily use.
2
4
891
《云端 AI 与隐私,我可以全都要吗?》最近调研了一下 Apple 在端侧模型和云端 AI 安全方面所做的 PCC 架构的努力,觉得很值得写博客分享出来让大家看到 Apple 针对这个场景结合 TEE 的 Solution,欢迎阅读:blog.cyyself.name/apple-pcc/
1
1
1
4
1,802